lustre/ptlrpc/events.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_RPC
  38
  39 #ifndef __KERNEL__
  40 # include <liblustre.h>
  41 #else
  42 # include <libcfs/libcfs.h>
  43 # ifdef __mips64__
  44 #  include <linux/kernel.h>
  45 # endif
  46 #endif
  47
  48 #include <obd_class.h>
  49 #include <lustre_net.h>
  50 #include <lustre_sec.h>
  51 #include "ptlrpc_internal.h"
  52
  53 lnet_handle_eq_t   ptlrpc_eq_h;
  54
  55 /*
  56  *  Client's outgoing request callback
  57  */
  58 void request_out_callback(lnet_event_t *ev)
  59 {
  60         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
  61         struct ptlrpc_request *req = cbid->cbid_arg;
  62         ENTRY;
  63
  64         LASSERT (ev->type == LNET_EVENT_SEND ||
  65                  ev->type == LNET_EVENT_UNLINK);
  66         LASSERT (ev->unlinked);
  67
  68         DEBUG_REQ((ev->status == 0) ? D_NET : D_ERROR, req,
  69                   "type %d, status %d", ev->type, ev->status);
  70
  71         sptlrpc_request_out_callback(req);
  72
  73         if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
  74
  75                 /* Failed send: make it seem like the reply timed out, just
  76                  * like failing sends in client.c does currently...  */
  77
  78                 spin_lock(&req->rq_lock);
  79                 req->rq_net_err = 1;
  80                 spin_unlock(&req->rq_lock);
  81
  82                 ptlrpc_client_wake_req(req);
  83         }
  84
  85         ptlrpc_req_finished(req);
  86
  87         EXIT;
  88 }
  89
  90 /*
  91  * Client's incoming reply callback
  92  */
  93 void reply_in_callback(lnet_event_t *ev)
  94 {
  95         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
  96         struct ptlrpc_request *req = cbid->cbid_arg;
  97         ENTRY;
  98
  99         DEBUG_REQ((ev->status == 0) ? D_NET : D_ERROR, req,
 100                   "type %d, status %d", ev->type, ev->status);
 101
 102         LASSERT (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
 103         LASSERT (ev->md.start == req->rq_repbuf);
 104         LASSERT (ev->mlength <= req->rq_repbuf_len);
 105         /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
 106            for adaptive timeouts' early reply. */
 107         LASSERT((ev->md.options & LNET_MD_MANAGE_REMOTE) != 0);
 108
 109         spin_lock(&req->rq_lock);
 110
 111         req->rq_receiving_reply = 0;
 112         req->rq_early = 0;
 113         if (ev->unlinked)
 114                 req->rq_must_unlink = 0;
 115
 116         if (ev->status)
 117                 goto out_wake;
 118         if (ev->type == LNET_EVENT_UNLINK) {
 119                 LASSERT(ev->unlinked);
 120                 DEBUG_REQ(D_RPCTRACE, req, "unlink");
 121                 goto out_wake;
 122         }
 123
 124         if ((ev->offset == 0) &&
 125             ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
 126                 /* Early reply */
 127                 DEBUG_REQ(D_ADAPTTO, req,
 128                           "Early reply received: mlen=%u offset=%d replen=%d "
 129                           "replied=%d unlinked=%d", ev->mlength, ev->offset,
 130                           req->rq_replen, req->rq_replied, ev->unlinked);
 131
 132                 req->rq_early_count++; /* number received, client side */
 133
 134                 if (req->rq_replied)   /* already got the real reply */
 135                         goto out_wake;
 136
 137                 req->rq_early = 1;
 138                 req->rq_reply_off = ev->offset;
 139                 req->rq_nob_received = ev->mlength;
 140                 /* And we're still receiving */
 141                 req->rq_receiving_reply = 1;
 142         } else {
 143                 /* Real reply */
 144                 req->rq_rep_swab_mask = 0;
 145                 req->rq_replied = 1;
 146                 req->rq_reply_off = ev->offset;
 147                 req->rq_nob_received = ev->mlength;
 148                 /* LNetMDUnlink can't be called under the LNET_LOCK,
 149                    so we must unlink in ptlrpc_unregister_reply */
 150                 DEBUG_REQ(D_INFO, req,
 151                           "reply in flags=%x mlen=%u offset=%d replen=%d",
 152                           lustre_msg_get_flags(req->rq_reqmsg),
 153                           ev->mlength, ev->offset, req->rq_replen);
 154         }
 155
 156         req->rq_import->imp_last_reply_time = cfs_time_current_sec();
 157
 158 out_wake:
 159         /* NB don't unlock till after wakeup; req can disappear under us
 160          * since we don't have our own ref */
 161         ptlrpc_client_wake_req(req);
 162         spin_unlock(&req->rq_lock);
 163         EXIT;
 164 }
 165
 166 /*
 167  * Client's bulk has been written/read
 168  */
 169 void client_bulk_callback (lnet_event_t *ev)
 170 {
 171         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 172         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 173         ENTRY;
 174
 175         LASSERT ((desc->bd_type == BULK_PUT_SINK &&
 176                   ev->type == LNET_EVENT_PUT) ||
 177                  (desc->bd_type == BULK_GET_SOURCE &&
 178                   ev->type == LNET_EVENT_GET) ||
 179                  ev->type == LNET_EVENT_UNLINK);
 180         LASSERT (ev->unlinked);
 181
 182         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 183                "event type %d, status %d, desc %p\n",
 184                ev->type, ev->status, desc);
 185
 186         spin_lock(&desc->bd_lock);
 187
 188         LASSERT(desc->bd_network_rw);
 189         desc->bd_network_rw = 0;
 190
 191         if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
 192                 desc->bd_success = 1;
 193                 desc->bd_nob_transferred = ev->mlength;
 194                 desc->bd_sender = ev->sender;
 195         }
 196
 197         sptlrpc_enc_pool_put_pages(desc);
 198
 199         /* NB don't unlock till after wakeup; desc can disappear under us
 200          * otherwise */
 201         ptlrpc_client_wake_req(desc->bd_req);
 202
 203         spin_unlock(&desc->bd_lock);
 204         EXIT;
 205 }
 206
 207 /*
 208  * Server's incoming request callback
 209  */
 210 void request_in_callback(lnet_event_t *ev)
 211 {
 212         struct ptlrpc_cb_id               *cbid = ev->md.user_ptr;
 213         struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
 214         struct ptlrpc_service             *service = rqbd->rqbd_service;
 215         struct ptlrpc_request             *req;
 216         ENTRY;
 217
 218         LASSERT (ev->type == LNET_EVENT_PUT ||
 219                  ev->type == LNET_EVENT_UNLINK);
 220         LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
 221         LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
 222                  rqbd->rqbd_buffer + service->srv_buf_size);
 223
 224         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 225                "event type %d, status %d, service %s\n",
 226                ev->type, ev->status, service->srv_name);
 227
 228         if (ev->unlinked) {
 229                 /* If this is the last request message to fit in the
 230                  * request buffer we can use the request object embedded in
 231                  * rqbd.  Note that if we failed to allocate a request,
 232                  * we'd have to re-post the rqbd, which we can't do in this
 233                  * context. */
 234                 req = &rqbd->rqbd_req;
 235                 memset(req, 0, sizeof (*req));
 236         } else {
 237                 LASSERT (ev->type == LNET_EVENT_PUT);
 238                 if (ev->status != 0) {
 239                         /* We moaned above already... */
 240                         return;
 241                 }
 242                 OBD_ALLOC_GFP(req, sizeof(*req), CFS_ALLOC_ATOMIC_TRY);
 243                 if (req == NULL) {
 244                         CERROR("Can't allocate incoming request descriptor: "
 245                                "Dropping %s RPC from %s\n",
 246                                service->srv_name,
 247                                libcfs_id2str(ev->initiator));
 248                         return;
 249                 }
 250         }
 251
 252         /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
 253          * flags are reset and scalars are zero.  We only set the message
 254          * size to non-zero if this was a successful receive. */
 255         req->rq_xid = ev->match_bits;
 256         req->rq_reqbuf = ev->md.start + ev->offset;
 257         if (ev->type == LNET_EVENT_PUT && ev->status == 0)
 258                 req->rq_reqdata_len = ev->mlength;
 259         do_gettimeofday(&req->rq_arrival_time);
 260         req->rq_peer = ev->initiator;
 261         req->rq_self = ev->target.nid;
 262         req->rq_rqbd = rqbd;
 263         req->rq_phase = RQ_PHASE_NEW;
 264 #ifdef CRAY_XT3
 265         req->rq_uid = ev->uid;
 266 #endif
 267         spin_lock_init(&req->rq_lock);
 268         CFS_INIT_LIST_HEAD(&req->rq_timed_list);
 269         atomic_set(&req->rq_refcount, 1);
 270         if (ev->type == LNET_EVENT_PUT)
 271                 DEBUG_REQ(D_RPCTRACE, req, "incoming req");
 272
 273         CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
 274
 275         spin_lock(&service->srv_lock);
 276
 277         req->rq_history_seq = service->srv_request_seq++;
 278         list_add_tail(&req->rq_history_list, &service->srv_request_history);
 279
 280         if (ev->unlinked) {
 281                 service->srv_nrqbd_receiving--;
 282                 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
 283                        service->srv_nrqbd_receiving);
 284
 285                 /* Normally, don't complain about 0 buffers posted; LNET won't
 286                  * drop incoming reqs since we set the portal lazy */
 287                 if (test_req_buffer_pressure &&
 288                     ev->type != LNET_EVENT_UNLINK &&
 289                     service->srv_nrqbd_receiving == 0)
 290                         CWARN("All %s request buffers busy\n",
 291                               service->srv_name);
 292
 293                 /* req takes over the network's ref on rqbd */
 294         } else {
 295                 /* req takes a ref on rqbd */
 296                 rqbd->rqbd_refcount++;
 297         }
 298
 299         list_add_tail(&req->rq_list, &service->srv_req_in_queue);
 300         service->srv_n_queued_reqs++;
 301
 302         /* NB everything can disappear under us once the request
 303          * has been queued and we unlock, so do the wake now... */
 304         cfs_waitq_signal(&service->srv_waitq);
 305
 306         spin_unlock(&service->srv_lock);
 307         EXIT;
 308 }
 309
 310 /*
 311  *  Server's outgoing reply callback
 312  */
 313 void reply_out_callback(lnet_event_t *ev)
 314 {
 315         struct ptlrpc_cb_id       *cbid = ev->md.user_ptr;
 316         struct ptlrpc_reply_state *rs = cbid->cbid_arg;
 317         struct ptlrpc_service     *svc = rs->rs_service;
 318         ENTRY;
 319
 320         LASSERT (ev->type == LNET_EVENT_SEND ||
 321                  ev->type == LNET_EVENT_ACK ||
 322                  ev->type == LNET_EVENT_UNLINK);
 323
 324         if (!rs->rs_difficult) {
 325                 /* 'Easy' replies have no further processing so I drop the
 326                  * net's ref on 'rs' */
 327                 LASSERT (ev->unlinked);
 328                 ptlrpc_rs_decref(rs);
 329                 atomic_dec (&svc->srv_outstanding_replies);
 330                 EXIT;
 331                 return;
 332         }
 333
 334         LASSERT (rs->rs_on_net);
 335
 336         if (ev->unlinked) {
 337                 /* Last network callback.  The net's ref on 'rs' stays put
 338                  * until ptlrpc_server_handle_reply() is done with it */
 339                 spin_lock(&svc->srv_lock);
 340                 spin_lock(&rs->rs_lock);
 341                 rs->rs_on_net = 0;
 342                 if (!rs->rs_no_ack ||
 343                     rs->rs_transno <= rs->rs_export->exp_obd->obd_last_committed)
 344                         ptlrpc_schedule_difficult_reply (rs);
 345                 spin_unlock(&rs->rs_lock);
 346                 spin_unlock(&svc->srv_lock);
 347         }
 348
 349         EXIT;
 350 }
 351
 352 /*
 353  * Server's bulk completion callback
 354  */
 355 void server_bulk_callback (lnet_event_t *ev)
 356 {
 357         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 358         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 359         ENTRY;
 360
 361         LASSERT (ev->type == LNET_EVENT_SEND ||
 362                  ev->type == LNET_EVENT_UNLINK ||
 363                  (desc->bd_type == BULK_PUT_SOURCE &&
 364                   ev->type == LNET_EVENT_ACK) ||
 365                  (desc->bd_type == BULK_GET_SINK &&
 366                   ev->type == LNET_EVENT_REPLY));
 367
 368         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 369                "event type %d, status %d, desc %p\n",
 370                ev->type, ev->status, desc);
 371
 372         spin_lock(&desc->bd_lock);
 373
 374         if ((ev->type == LNET_EVENT_ACK ||
 375              ev->type == LNET_EVENT_REPLY) &&
 376             ev->status == 0) {
 377                 /* We heard back from the peer, so even if we get this
 378                  * before the SENT event (oh yes we can), we know we
 379                  * read/wrote the peer buffer and how much... */
 380                 desc->bd_success = 1;
 381                 desc->bd_nob_transferred = ev->mlength;
 382                 desc->bd_sender = ev->sender;
 383         }
 384
 385         if (ev->unlinked) {
 386                 /* This is the last callback no matter what... */
 387                 desc->bd_network_rw = 0;
 388                 cfs_waitq_signal(&desc->bd_waitq);
 389         }
 390
 391         spin_unlock(&desc->bd_lock);
 392         EXIT;
 393 }
 394
 395 static void ptlrpc_master_callback(lnet_event_t *ev)
 396 {
 397         struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
 398         void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
 399
 400         /* Honestly, it's best to find out early. */
 401         LASSERT (cbid->cbid_arg != LP_POISON);
 402         LASSERT (callback == request_out_callback ||
 403                  callback == reply_in_callback ||
 404                  callback == client_bulk_callback ||
 405                  callback == request_in_callback ||
 406                  callback == reply_out_callback ||
 407                  callback == server_bulk_callback);
 408
 409         callback (ev);
 410 }
 411
 412 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
 413                          lnet_process_id_t *peer, lnet_nid_t *self)
 414 {
 415         int               best_dist = 0;
 416         __u32             best_order = 0;
 417         int               count = 0;
 418         int               rc = -ENOENT;
 419         int               portals_compatibility;
 420         int               dist;
 421         __u32             order;
 422         lnet_nid_t        dst_nid;
 423         lnet_nid_t        src_nid;
 424
 425         portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
 426
 427         peer->pid = LUSTRE_SRV_LNET_PID;
 428
 429         /* Choose the matching UUID that's closest */
 430         while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
 431                 dist = LNetDist(dst_nid, &src_nid, &order);
 432                 if (dist < 0)
 433                         continue;
 434
 435                 if (dist == 0) {                /* local! use loopback LND */
 436                         peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
 437                         rc = 0;
 438                         break;
 439                 }
 440
 441                 if (rc < 0 ||
 442                     dist < best_dist ||
 443                     (dist == best_dist && order < best_order)) {
 444                         best_dist = dist;
 445                         best_order = order;
 446
 447                         if (portals_compatibility > 1) {
 448                                 /* Strong portals compatibility: Zero the nid's
 449                                  * NET, so if I'm reading new config logs, or
 450                                  * getting configured by (new) lconf I can
 451                                  * still talk to old servers. */
 452                                 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
 453                                 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
 454                         }
 455                         peer->nid = dst_nid;
 456                         *self = src_nid;
 457                         rc = 0;
 458                 }
 459         }
 460
 461         CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
 462         if (rc != 0)
 463                 CERROR("No NID found for %s\n", uuid->uuid);
 464         return rc;
 465 }
 466
 467 void ptlrpc_ni_fini(void)
 468 {
 469         cfs_waitq_t         waitq;
 470         struct l_wait_info  lwi;
 471         int                 rc;
 472         int                 retries;
 473
 474         /* Wait for the event queue to become idle since there may still be
 475          * messages in flight with pending events (i.e. the fire-and-forget
 476          * messages == client requests and "non-difficult" server
 477          * replies */
 478
 479         for (retries = 0;; retries++) {
 480                 rc = LNetEQFree(ptlrpc_eq_h);
 481                 switch (rc) {
 482                 default:
 483                         LBUG();
 484
 485                 case 0:
 486                         LNetNIFini();
 487                         return;
 488
 489                 case -EBUSY:
 490                         if (retries != 0)
 491                                 CWARN("Event queue still busy\n");
 492
 493                         /* Wait for a bit */
 494                         cfs_waitq_init(&waitq);
 495                         lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
 496                         l_wait_event(waitq, 0, &lwi);
 497                         break;
 498                 }
 499         }
 500         /* notreached */
 501 }
 502
 503 lnet_pid_t ptl_get_pid(void)
 504 {
 505         lnet_pid_t        pid;
 506
 507 #ifndef  __KERNEL__
 508         pid = getpid();
 509 #else
 510         pid = LUSTRE_SRV_LNET_PID;
 511 #endif
 512         return pid;
 513 }
 514
 515 int ptlrpc_ni_init(void)
 516 {
 517         int              rc;
 518         lnet_pid_t       pid;
 519
 520         pid = ptl_get_pid();
 521         CDEBUG(D_NET, "My pid is: %x\n", pid);
 522
 523         /* We're not passing any limits yet... */
 524         rc = LNetNIInit(pid);
 525         if (rc < 0) {
 526                 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
 527                 return (-ENOENT);
 528         }
 529
 530         /* CAVEAT EMPTOR: how we process portals events is _radically_
 531          * different depending on... */
 532 #ifdef __KERNEL__
 533         /* kernel portals calls our master callback when events are added to
 534          * the event queue.  In fact lustre never pulls events off this queue,
 535          * so it's only sized for some debug history. */
 536         rc = LNetEQAlloc(1024, ptlrpc_master_callback, &ptlrpc_eq_h);
 537 #else
 538         /* liblustre calls the master callback when it removes events from the
 539          * event queue.  The event queue has to be big enough not to drop
 540          * anything */
 541         rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &ptlrpc_eq_h);
 542 #endif
 543         if (rc == 0)
 544                 return 0;
 545
 546         CERROR ("Failed to allocate event queue: %d\n", rc);
 547         LNetNIFini();
 548
 549         return (-ENOMEM);
 550 }
 551
 552 #ifndef __KERNEL__
 553 CFS_LIST_HEAD(liblustre_wait_callbacks);
 554 CFS_LIST_HEAD(liblustre_idle_callbacks);
 555 void *liblustre_services_callback;
 556
 557 void *
 558 liblustre_register_waitidle_callback (struct list_head *callback_list,
 559                                       const char *name,
 560                                       int (*fn)(void *arg), void *arg)
 561 {
 562         struct liblustre_wait_callback *llwc;
 563
 564         OBD_ALLOC(llwc, sizeof(*llwc));
 565         LASSERT (llwc != NULL);
 566
 567         llwc->llwc_name = name;
 568         llwc->llwc_fn = fn;
 569         llwc->llwc_arg = arg;
 570         list_add_tail(&llwc->llwc_list, callback_list);
 571
 572         return (llwc);
 573 }
 574
 575 void
 576 liblustre_deregister_waitidle_callback (void *opaque)
 577 {
 578         struct liblustre_wait_callback *llwc = opaque;
 579
 580         list_del(&llwc->llwc_list);
 581         OBD_FREE(llwc, sizeof(*llwc));
 582 }
 583
 584 void *
 585 liblustre_register_wait_callback (const char *name,
 586                                   int (*fn)(void *arg), void *arg)
 587 {
 588         return liblustre_register_waitidle_callback(&liblustre_wait_callbacks,
 589                                                     name, fn, arg);
 590 }
 591
 592 void
 593 liblustre_deregister_wait_callback (void *opaque)
 594 {
 595         liblustre_deregister_waitidle_callback(opaque);
 596 }
 597
 598 void *
 599 liblustre_register_idle_callback (const char *name,
 600                                   int (*fn)(void *arg), void *arg)
 601 {
 602         return liblustre_register_waitidle_callback(&liblustre_idle_callbacks,
 603                                                     name, fn, arg);
 604 }
 605
 606 void
 607 liblustre_deregister_idle_callback (void *opaque)
 608 {
 609         liblustre_deregister_waitidle_callback(opaque);
 610 }
 611
 612 int
 613 liblustre_check_events (int timeout)
 614 {
 615         lnet_event_t ev;
 616         int         rc;
 617         int         i;
 618         ENTRY;
 619
 620         rc = LNetEQPoll(&ptlrpc_eq_h, 1, timeout * 1000, &ev, &i);
 621         if (rc == 0)
 622                 RETURN(0);
 623
 624         LASSERT (rc == -EOVERFLOW || rc == 1);
 625
 626         /* liblustre: no asynch callback so we can't affort to miss any
 627          * events... */
 628         if (rc == -EOVERFLOW) {
 629                 CERROR ("Dropped an event!!!\n");
 630                 abort();
 631         }
 632
 633         ptlrpc_master_callback (&ev);
 634         RETURN(1);
 635 }
 636
 637 int liblustre_waiting = 0;
 638
 639 int
 640 liblustre_wait_event (int timeout)
 641 {
 642         struct list_head               *tmp;
 643         struct liblustre_wait_callback *llwc;
 644         int                             found_something = 0;
 645
 646         /* single threaded recursion check... */
 647         liblustre_waiting = 1;
 648
 649         for (;;) {
 650                 /* Deal with all pending events */
 651                 while (liblustre_check_events(0))
 652                         found_something = 1;
 653
 654                 /* Give all registered callbacks a bite at the cherry */
 655                 list_for_each(tmp, &liblustre_wait_callbacks) {
 656                         llwc = list_entry(tmp, struct liblustre_wait_callback,
 657                                           llwc_list);
 658
 659                         if (llwc->llwc_fn(llwc->llwc_arg))
 660                                 found_something = 1;
 661                 }
 662
 663                 if (found_something || timeout == 0)
 664                         break;
 665
 666                 /* Nothing so far, but I'm allowed to block... */
 667                 found_something = liblustre_check_events(timeout);
 668                 if (!found_something)           /* still nothing */
 669                         break;                  /* I timed out */
 670         }
 671
 672         liblustre_waiting = 0;
 673
 674         return found_something;
 675 }
 676
 677 void
 678 liblustre_wait_idle(void)
 679 {
 680         static int recursed = 0;
 681
 682         struct list_head               *tmp;
 683         struct liblustre_wait_callback *llwc;
 684         int                             idle = 0;
 685
 686         LASSERT(!recursed);
 687         recursed = 1;
 688
 689         do {
 690                 liblustre_wait_event(0);
 691
 692                 idle = 1;
 693
 694                 list_for_each(tmp, &liblustre_idle_callbacks) {
 695                         llwc = list_entry(tmp, struct liblustre_wait_callback,
 696                                           llwc_list);
 697
 698                         if (!llwc->llwc_fn(llwc->llwc_arg)) {
 699                                 idle = 0;
 700                                 break;
 701                         }
 702                 }
 703
 704         } while (!idle);
 705
 706         recursed = 0;
 707 }
 708
 709 #endif /* __KERNEL__ */
 710
 711 int ptlrpc_init_portals(void)
 712 {
 713         int   rc = ptlrpc_ni_init();
 714
 715         if (rc != 0) {
 716                 CERROR("network initialisation failed\n");
 717                 return -EIO;
 718         }
 719 #ifndef __KERNEL__
 720         liblustre_services_callback =
 721                 liblustre_register_wait_callback("liblustre_check_services",
 722                                                  &liblustre_check_services,
 723                                                  NULL);
 724         init_completion_module(liblustre_wait_event);
 725 #endif
 726         rc = ptlrpcd_addref();
 727         if (rc == 0)
 728                 return 0;
 729
 730         CERROR("rpcd initialisation failed\n");
 731 #ifndef __KERNEL__
 732         liblustre_deregister_wait_callback(liblustre_services_callback);
 733 #endif
 734         ptlrpc_ni_fini();
 735         return rc;
 736 }
 737
 738 void ptlrpc_exit_portals(void)
 739 {
 740 #ifndef __KERNEL__
 741         liblustre_deregister_wait_callback(liblustre_services_callback);
 742 #endif
 743         ptlrpcd_decref();
 744         ptlrpc_ni_fini();
 745 }